Fig. 4.

Predictive performance assessment. (A) Scatterplot of the performance of PRECISE and Ridge regression. Predictive performance for each approach was computed as the Pearson correlation between the measured and the 10-fold cross-validated predicted ${\text{IC}}_{50}$s. PRECISE and Ridge regression performance is strongly correlated, with PRECISE showing a slight drop in performance. (B) Predicted drug response (predicted ${\text{IC}}_{50}$) for breast tumors based on a predictor of Lapatinib response trained on all the cell lines—i.e. across all tissue types—employing gene expression data only. ERBB2 copy number status in tumors correlates significantly with predicted ${\text{IC}}_{50}$ values, validating the predictions, as tumors with ERBB2 amplifications are known to be more sensitive to Lapatinib. (C) Predicted drug response (predicted ${\text{IC}}_{50}$) for skin melanoma tumors based on a predictor of Tramatinib response trained on all the cell lines—i.e. across all tissue types—employing gene expression data only. The PRECISE predictions are validated by the established fact that ${\text{BRAF}}^{\mathrm{V}600\mathrm{E}}$-mutated tumors show a significantly higher sensitivity while the tumors bearing other mutations in $\text{BRAF}$ do not show responses that differ from wild-type tumors